Hyperspectral image classification is challenging due to high spectral dimensionality, redundancy, acquisition noise, and spatial-spectral variability. This paper presents GeoSpectra, a noise-aware machine learning framework for hyperspectral mineral and land-cover classification using the Indian Pines dataset. The proposed framework applies band-wise Gaussian spatial smoothing with σ = 0.5, background removal, spectral standardization, and Principal Component Analysis (PCA) retaining 99% of the variance. Support Vector Machine (SVM) and Random Forest (RF) classifiers are optimized using grid-search cross-validation and evaluated using a 70:30 training-testing split. The SVM achieves 91.61% accuracy, outperforming Random Forest at 84.20%. For application-level inference, the framework generates three spectral representations: raw, Fourier-filtered, and Fourier-plus-CNN reconstructed data. A one-dimensional CNN autoencoder is used for spectral reconstruction, while Fourier filtering suppresses structured stripe artifacts. A multi-path inference safeguard evaluates the representations using predicted-class diversity and PCA feature spread to identify collapsed or less informative outputs. In diagnostic evaluation, the Fourier-only path collapsed to a single predicted class and was rejected, while the raw path retained predictions across all 16 classes. The safeguard is designed as an inference-quality mechanism rather than as a demonstrated accuracy-improvement method. The complete framework is deployed as a Flask web application supporting authenticated uploads, classifier selection, hyperspectral classification-map visualization, and pixel-wise prediction downloads. The proposed system integrates preprocessing, dimensionality reduction, machine learning classification, denoising, inference-quality checking, and web deployment into a unified hyperspectral analysis framework.